Remaining Oil Distribution Prediction Using LSTM and Kriging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting remaining oil distribution require significant data preprocessing and separate calculations for each grid cell, leading to low prediction efficiency.
Innovation Solution
A method combining historical and predictive reservoir knowledge with deep learning technology and traditional interpolation methods, using LSTM models and Kriging interpolation, to efficiently and accurately predict remaining oil distribution by establishing complete data sets, setting boundary conditions, and deriving oil-gas-water three-phase saturation fields.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If artificial intelligence methods are used to calculate remaining oil distribution by gridding the reservoir and establishing SVM classification models and deep learning models for each grid cell, then prediction accuracy is improved, but data preprocessing workload increases significantly and prediction efficiency decreases
Solution Approach 1:
The patent divides the reservoir into flow units based on heterogeneity characteristics, and then performs separate remaining oil distribution predictions for each flow unit. This segmentation approach reduces the complexity of the overall problem while maintaining prediction accuracy within each unit, avoiding the need to process the entire reservoir as a single grid system.
Solution Approach 2:
The patent establishes a unified remaining oil distribution prediction model that can be applied across different flow units and reservoir conditions. The model uses universal input parameters (production data, injection data, well test data) and produces consistent output formats, enabling efficient prediction across the entire reservoir without requiring separate models for each grid cell.
2Manufacturing precision
If artificial intelligence methods grid the reservoir and calculate each grid cell separately, then detailed spatial distribution is achieved, but computational complexity and preprocessing requirements increase
Solution Approach 1:
The reservoir is segmented into flow units based on heterogeneity characteristics, and remaining oil distribution is predicted separately for each flow unit. This approach captures spatial distribution details within each unit while avoiding the computational burden of grid-cell-by-grid-cell calculations across the entire reservoir.
Solution Approach 2:
The patent introduces flow units as intermediate entities between the reservoir-scale model and grid-cell-level details. These flow units serve as mediators that aggregate spatial information at an appropriate scale, providing detailed spatial distribution where needed while reducing overall computational complexity through hierarchical modeling.
Data Source
AI summary
The present disclosure provides a method and device for predicting remaining oil distribution based on historical and predictive reservoir knowledge. The method includes: establishing complete data sets according to relevant historical dynamic and static monitor data of oil wells in a target reservoir; training and testing a long and short-term memory (LSTM) model using the complete data sets; predicting production of a single well at a preset moment by using the trained LSTM model; acquiring three-phase saturations of the single well at the preset moment based on the production predicting results and experience of reservoir experts; and according to the three-phase saturations of the single well, setting boundary conditions and physical constraints for the reservoir, and deriving an oil-gas-water three-phase saturation field at the preset moment using Kriging interpolation to obtain a prediction result of remaining oil distribution in the target reservoir.